Oncology Ventures Thesis: Artificial Intelligence in Cancer Care

The oncology world is experiencing rapid growth in data that are collected to enhance cancer care. Through improvements in artificial intelligence (AI) and computing hardware advances, there is now a computational basis to integrate and synthesize this quickly expanding set of multi-dimensional data. 

Oncology is particularly poised for transformative changes brought on by AI, given the proven advantages of personalized care (e.g. we know that tumors and their response rates differ vastly from person to person).  

---

Current State of AI in Cancer

AI is increasingly being used in cancer research and to make treatment decisions. But, it is not that easy to deploy AI into the cancer care continuum at scale.

Effective AI support relies on the incorporation of disparate, complex data streams, including clinical presentation, patient history, tumor pathology and genomic profiles, as well as medical imaging, all while matching these data to the findings of an ever-growing body of scientific literature. 

We have a long way to go to ensure that AI can be deployed and commercialized in an oncology clinical workflow. Key issues include:

• Data quality: Medical data, as we know, is often messy and incomplete. This can lead to errors and bias in the AI algorithms.

• Data sharing: I don’t know if you have noticed this in the healthcare world, but sharing medical data across institutions can be challenging. Building large, shareable datasets is crucial for the development of AI algorithms that are robust and generalizable.

• Explainability: AI models can be difficult to interpret and explain, which can limit their trustworthiness and acceptance by clinicians and patients. If it isn’t easily understood, it is tough to validate and scale the model. 

• Generalization: An AI model that is developed using one type of cancer or patient population may not be applicable to other types of cancer or patient populations.

• Integration into the healthcare workflow: AI models need to be integrated into existing clinical systems in a way that is seamless and efficient.

• Regulatory approval: Developing and deploying AI-based medical devices and treatments requires regulatory approval, which is complex and time-consuming.

While we are at it, the workflow for an AI cancer researcher to publish a model could use some work. Here is what the workflow looks like today: 

• Data collection: The researcher collects and curates a dataset of medical records, images, and/or genomic data related to the cancer type of interest.

  • This data is typically difficult to search for within their own institution, let alone try to find from another healthcare system.

• Model development: They develop their model, typically on a platform like GitHub, and seek to analyze the data and make predictions about cancer diagnoses, treatment, or prognoses.

  • The issues here are that 1) this is being done in a siloed setting and 2) these platforms don’t overlay important meta data that is required in healthcare, such as SNOMED ontology, or allow for FHIR and EHR integrations.

• Model evaluation: The researcher evaluates the performance of the AI model using various metrics, such as accuracy and precision.

  • There is a rinse and repeat process that happens from here that could use a lot of help on the automation front.

• Manuscript preparation: The researcher prepares a manuscript describing the AI model, including details on the dataset, model architecture, evaluation metrics, and validation results.

  • This is done on outdated infrastructure, like Microsoft Word, and standardization is difficult to achieve.

• Peer review: The manuscript is submitted to a peer-reviewed scientific journal, where it undergoes a rigorous review process by independent experts in the field.

  • Given how slow, siloed and outdated that infrastructure is, the process takes months instead of hours.

• Publication: If the manuscript is accepted for publication, it is published in the scientific journal and becomes part of the scientific literature.

• Deployment: The AI model may be deployed in clinical settings or incorporated into clinical decision support systems for use by healthcare providers.

  • Here you have to deal with the likes of your academic medical center’s tech transfer office, which is time consuming, expensive and typically results in your model (which we have proven works) sitting on the shelf. 

---

Where is the Opportunity for AI to impact Oncology? 

Oncology is the world’s largest pharmaceutical therapeutic area. Over the past five years, ~$6 billion has been invested in companies engaged in the development of AI in oncology-based software solutions.

We believe that at this point in time, given the challenges outlined above, the most successful models will leverage large-scale, robustly annotated datasets to help complete narrow tasks at specific cancer care touchpoints. Below is a great example of a cancer patient pathway, and how AI can assist different users at different stages of the treatment journey, with a quickly expanding set of data. 

Source: Artificial intelligence for clinical oncology

Some examples of AI applications in oncology include:

• Drug discovery/development: 

  • Oncology makes up nearly 40 percent of the global clinical pipeline.

  • Identifying effective drug candidates involves screening large libraries of compounds to see which have the optimal biological activity with the fewest side effects. 

  • Machine learning algorithms can be trained on chemical compounds to screen these libraries much faster and more accurately, getting to a valid drug candidate faster.

• Clinical trial matching: 

  • Currently, analyzing data from multiple sources to identify eligible participants in clinical trials is difficult and requires a lot of manual effort.

  • Machine Learning (ML) and Natural Language Processing (NLP) algorithms can scan large datasets of patients to identify participants that are most likely to respond to treatment much faster, ultimately getting drugs to market quicker.

• Personalized medicine: 

  • Within the oncology drug pipeline, 70% are targeted therapeutics. Precision medicine is becoming increasingly vital to enhance oncology care.

  • Integrating data from genomic data, clinical data, imaging, etc. to make a tailored treatment plan can be challenging. Then you have to interpret the data to identify the important mutation or biomarker.

  • Now, a combination of ML and NLP can be used to pull together all of a patient’s information to identify a biomarker that will connect them with the treatment they are most likely to respond to.

• Cancer diagnosis: 

  • AI algorithms are being developed to improve the accuracy of cancer diagnosis through analysis of medical imaging such as CT scans, MRI and X-rays. 

  • For example, AI-powered algorithms have been shown to detect lung cancer in medical imaging with high accuracy.

• Incentivizing focus and research for rare cancers:

  • Manufacturers could choose to target large groups of less common cancers, those with fewer than 5,000 patients per year. 

  • While unmet need is very high for some indications, such as glioblastoma (an aggressive type of brain cancer), smaller patient numbers require a new approach to develop and commercialize effective therapies, which the right AI models can help to support and effectively scale.

What needs to happen to effectively implement this opportunity?

Developing appropriate reimbursement models for AI-driven care services is crucial. 

Insurers and other payors need to quickly recognize the value of AI and be willing to cover these services. Incentive structures need to change to encourage cost-effective, preventive and personalized care driven by AI, while ensuring that providers are fairly compensated for leveraging that technology. Of course, regulatory alignment here is key. 

Then, we need user-adoption. Clinicians should be encouraged to adopt AI tools and integrate them into their practice. We have to set up a structure to train and support practitioners so they can understand how AI can enhance their decision-making, reduce diagnostic errors and streamline care processes. 

Designing solutions that are as unobtrusive as possible for physicians and both trustworthy and acceptable by patients should be a core focus area. Incentives should follow suit, focused on improving patient outcomes and clinician satisfaction.

Oncology Ventures is excited to partner with and support the next generation of AI start-ups making cancer care more affordable, efficient and accessible.